A Lightweight Deep Learning Model for Aerial Image classification and Its Application to UAV-based Disaster Management
Deng Xinjie, Burhan Khan, Fazal Ghaffar, Yit Hong Choo, Tao Zhou · Procedia Computer Science · 2024
Unmanned Aerial Vehicles (UAVs) play a pivotal role in disaster management and emergency response for ensuring safety of our society. In this respect, lightweight deep learning models are necessary for video and image processing onboard of UAVs for real-time applications. In this study, a useful feature extraction module known as “multi-block” is designed to effectively capture information across multiple imagery scales for facilitate lightweight deep learning models. The multi-block module incorporates depthwise convolution layers with diverse kernel sizes to ensure robust feature extraction. Then, pointwise convolution is leveraged to capture the correlation between different spatial locations and channels. Evaluated with the Aerial Image Database for Emergency Response (AIDER), the proposed multi-block module demonstrates a reliable performance pertaining to its feature extraction capabilities for undertaking aerial image classification tasks in disaster management.